Introduction

Pituitary adenomas are among the most frequent intracranial tumors, with a 1/1500 prevalence [1]. They are mostly benign and typically presents with hormone hypersecretion syndromes and/or mass effect signs. Small intrasellar tumors can be clinically silent and diagnosed only as incidental MRI findings. Nonetheless, some entities among the different subtypes of adenoma show more aggressive and unpredictable behavior: these—such as sparsely granulated somatotroph adenomas, Crooke cell adenomas, and plurihormonal Pit-1-positive adenomas—tend to disclose local invasiveness and high-recurrence risk and rarely present features of highly aggressive cancer. Surgery is considered the first line of treatment in most cases but in this latter group, multimodality management therapy—as per neuro-oncological guidelines—is mandatory. The possibility of predicting pituitary tumor behavior at the preoperative stage cannot be yet considered reliable as no valid factor has been identified, and though it remains the critical groundwork of the pathologists.

The 2017 WHO classification revealed that former categories of typical adenomas, atypical adenomas, and pituitary carcinomas have no relevance from a clinical standpoint, introducing the definition of “high-risk” adenomas in reference to tumors with rapid growth, radiological invasion, and high Ki-67 proliferation index [2,3,4].

Considering the above, early identification of any radiological feature defining such behavior can be crucial in order to allow timely diagnosis and treatment. In this regard, radiomics, consisting of conversion of images into mineable data and subsequent analysis for decision support, has been emerging [5]. In particular, texture analysis is a post-processing technique for quantitative parameter extraction from pixel gray level heterogeneity. More recently, texture analysis-derived features have been used in association with data mining machine learning algorithms, aiding in the interpretation of a large amount of information produced. Machine learning is the branch of artificial intelligence dealing with computer algorithms capable of learning and improving in accuracy by analyzing datasets, without prior explicit programming [6]. It leads to the creation of automated predictive models to solve classification problems. The usefulness of the radiomics approach is being assessed in different fields of radiology [7,8,9,10,11,12,13,14].

The aim of our study was to assess the accuracy of machine learning analysis of texture-derived parameters from preoperative MRI of pituitary macroadenomas for the prediction of ki-67 proliferation index class.

Materials and methods

This study was approved by the local institutional review board, which waived the necessity for informed consent due to the retrospective nature of the study.

Subjects

We retrospectively reviewed data of 108 consecutive patients, who underwent endoscopic endonasal procedures for pituitary adenoma removal between January 2013 and December 2017, at the University of Naples “Federico II” Neurosurgery Unit.

Only patients with available ki-67 labeling index in the histopathological report were included (n = 106). Exclusion criteria were as follows: any previous treatment for pituitary adenoma (radiation or medical therapy) (n = 9), extensively necrotic or hemorrhagic lesions (n = 6), significant artifacts on the images used for the analysis (n = 2).

Demographic data, preoperative assessment, tumor features, and histopathological characteristics were retrieved from our electronic database (Filemaker Pro 11—File Maker Inc., Santa Clara, CA, USA).

Surgical approach

Endonasal surgical procedures were performed using a rigid 0-degree endoscope, 18 cm in length, and 4 mm in diameter (Karl Storz Endoscopy, Tuttlingen, Germany), as the sole visualizing tool. The use of 30–45° angled endoscopes was reserved to explore large intrasuprasellar post-surgical tumor cavities. They were run according to techniques already described in previous publications [15,16,17,18,19].

Pathological data

Specimens were obtained as formalin-fixed tissue. For the evaluation of the proliferation index ki-67 (labeling index (LI)), “hotspot” areas were chosen at low magnification and an average of the values on 5 adjacent fields (at least 500 neoplastic cells) was calculated (Fig. 1).

Fig. 1
figure 1

Segmentation examples on coronal T2-weighted images in two patients with low (upper row, a) and high (lower row, d) proliferation index pituitary adenomas, showing hand-drawn ROI placement (b and e, respectively). Pictures c and f show corresponding immunohistochemical evaluation (× 40 magnification) of ki-67 cell labeling index, respectively, 1 and 6%

Full size image

MRI acquisition data

Imaging studies were performed either on a 1.5-Tesla scanner (Gyroscan Intera, Philips, Eindhoven, The Netherlands) or 3T MR scanner (Magnetom Trio, Siemens Medical Solutions, Erlangen, Germany). Both protocols included a coronal T2-w TSE sequence (TR/TE, 2600/89 ms; FOV, 180 × 180 mm; matrix, 288 × 288; thk, 3 mm, at 1.5 T; TR/TE, 3000/98 ms FOV, 200 × 200 mm; matrix, 384 × 384; thk, 3 mm, at 3 T).

Image analysis

Tumors were identified by an expert neuroradiologist who proceeded to manual segmentation using a bidimensional polygonal ROI on the slice of lesion maximum extension with further editing with a brush tool, when needed. This process was carried on using freely available segmentation software (ITKSnap v3.6.0) [20] (Fig. 1). A second segmentation on a sample of 35 patients was performed by another neuroradiologist, blinded to the prior annotations and clinical data, to assess inter-operator reproducibility of the selected features.

Image pre-processing and feature extraction were performed in an open-source Python radiomics software (Pyradiomics v2.1.2) [21]. In particular, as two MRI scanners by different vendors and with distinct acquisition parameters were employed, image voxels were first normalized by subtracting the mean intensity and dividing by the standard deviation with an expected resulting range [− 3, 3], a mean of 0 and standard deviation of 1 in the normalized image. Then, a scale of 100 was applied, resulting in an intensity range mainly within [− 300, 300]. This would have been required even using a single scanner as non-quantitative sequences were evaluated (e.g., T2-w images and not T2 maps). All images and corresponding ROI masks were also resampled to a 2 × 2 × 2 mm resolution. To optimize extraction time, making it less memory intensive, images were automatically cropped based on ROI masks keeping 10-voxel padding in order to consent subsequent filter application.

As for image discretization, the employed software allows both for a fixed bin count and bin width. As suggested by the developers, we decided to use a fixed bin width that allowed us to obtain an ideal bin count between 16 and 128 [22]. To do so, a preliminary extraction of first order parameters was performed in all patients, calculating using the gray level range an optimal bin width of 3. To avoid negative gray values after normalization, potentially affecting first-order feature extraction, an array shift of 300 (3 standard deviations × 100) was applied.

Various filters were also used to generate derived images for additional texture feature extraction. In particular, wavelet decomposition yields all possible combinations of high- and low-pass filtering in the x and y dimensions; edge enhancement Laplacian of Gaussian (LoG) filter emphasizing gray level changes at different texture coarseness. Due to the applied resampling resolution and lesion size, we chose four LoG sigma values ranging from 2.0 (maximum fineness) to 3.5 mm (maximum coarseness) with 0.5-mm increments.

Regarding feature extraction, to bidimensional shape and first-order statistics, commonly referred to as histogram analysis, we added higher order class parameters. In detail, the symmetrical gray level co-occurrence matrix (GLCM) characterizes image texture by calculating how often pairs of voxels with specific intensity levels and spatial relationship occur in an image, extracting statistical measures from the deriving matrix [23]. Gray level size zone matrix (GLSZM) features quantify gray level zones, defined as the number of connected voxels that share the same intensity value [24]. The gray level run length matrix (GLRLM) evaluates gray level runs, which are the length of consecutive pixels with the same gray level [25]. Neighboring gray-tone difference matrix (NGTDM) features assess differences between pixel values and neighbor average gray value within a set distance [26]. Finally, employing gray level dependence matrix (GLDM)-derived parameters, the dependency of connected voxels, expressed as their number within a set distance, to the center voxel can be determined [27].

Statistical analysis

Feature scaling was performed through normalization. Subsequently, in order to remove irrelevant and redundant data, which can reduce computation time, improve learning accuracy, and facilitate a better understanding for the learning model or data, different supervised feature selection methods were employed [28]. This process and subsequent steps in our machine learning pipeline were performed on a freely available data mining software (Weka v3.8) [29]. To balance data dimensionality reduction and loss of information, we tested eight-feature selection method belonging to three different families as follows:

  1. 1.

    Embedded classifier methods, both ranking attribute worth with the OneR (1R) algorithm and C4.5 decision tree, and selecting the best feature subset for C4.5 decision tree and random tree algorithms [30];

  2. 2.

    Feature ranking methods using a direct Pearson’s correlation and a ReliefF Evaluator, a filter-method approach sensitive to feature interactions [31];

  3. 3.

    Subset feature search selection method, which evaluates attribute subset worth considering both feature predictive ability and redundancy [31].

Absolute agreement intraclass correlation coefficient (ICC) values were calculated for the selected texture features using SPSS version 17.

A k-nearest neighbors (k-NN) classifier was employed to predict macroadenoma proliferation index class. It represents a non-parametric, lazy learning algorithm, not making any assumption on the underlying data distribution and not using training data points to do generalization, deferring computation until classification [32]. We utilized a k = 3 linear nearest neighbor search with Euclidean distance function and no distance weighting.

Algorithm validation was performed with a train-test approach, randomly splitting the patients in training (60%) and test (40%) groups. Feature selection and model training were performed on the first while the latter was employed for classifier validation and calculation of accuracy metrics.

Our radiomics workflow pipeline is illustrated in Fig. 2.

Fig. 2
figure 2

Radiomic workflow pipeline. Manual segmentation of the lesions was performed on coronal T2-weighted images using a bidimensional polygonal ROI. Image pre-processing included resampling of all images and corresponding ROI masks to a 2 × 2 × 2 mm resolution, discretization, and normalization of voxel gray levels. Feature extraction (histogram and texture analysis) was performed on both native and filtered images. Subsequently, in order to remove irrelevant and redundant data, different supervised feature selection methods were employed. A k-NN classifier was employed to predict the macroadenoma proliferation index class with a train-test approach

Full size image

Results

Subjects and pathological data

According to the inclusion and exclusion criteria, the final population consisted of 89 patients (51 males and 38 females; mean age 52.17 ± 1 years; range, 16–80). As stated above, they were divided into training (n = 53) and test (n = 36) groups.

Of the included lesions, 25 were functioning adenomas (5 ACTH, 8 GH, 5 GH/PRL, 6 PRL, and 1 TSH secreting) and 64 non-functioning. Concerning tumor location, 13 were purely intrasellar, 40 presented suprasellar infradiaphragmatic extension, while 36 involved the supradiaphragmatic space (Table 1).

Table 1 Patient population clinical data
Full size table

Surgical treatment was performed via “standard” endoscopic endonasal approach in 79 cases, with a transtuberculum transplanum “extended” approach in the remaining. Overall gross-total removal was obtained in 74 cases with a subtotal removal in the other 15 patients.

In regard to pathological features, we identified 59 lesions (12 functioning and 47 non-functioning tumors) whose ki-67 was lesser than 3% (low proliferation index), while in 30 cases (13 functioning and 17 non-functioning lesions), it resulted equal to or greater than 3% (high proliferative index) [33].

The training group included 20 high proliferation and 33 low proliferation index patients, while the test one included respectively 10 and 26 patients for each class.

Image analysis

A total of 1128 features were extracted from each patient, represented on the correlation cluster map shown in Fig. 3. Of the 12 subsets derived from feature selection (feature number range = 1–6), the best performing one was constituted by the 4 highest correlating parameters at Pearson’s test. They were the following: kurtosis derived from the filtered LoG (sigma = 3.5), zone variance from both the GLSZM of the original image and after low-high pass wavelet transform, and large area emphasis from the GLSZM after low-high pass wavelet transform. Their ICC values were respectively 0.87, 0.97, 0.93, and 0.95, indicating very good reproducibility [34]. Figures 4 and 5 present univariate and pairwise feature distribution for this data subset in our population.

Fig. 3
figure 3

Feature correlation matrix represented as a hierarchically clustered heatmap

Full size image
Fig. 4
figure 4

Box plot of the distribution for each feature of the selected subset in relation to the proliferation index class

Full size image
Fig. 5
figure 5

Pairwise bivariate distribution with regression lines for the selected feature subset in relation to the proliferation index class

Full size image

The k-NN overall accuracy was 91.67% (33/36) of correctly classified patients. Other evaluation metrics are reported in Table 2, derived from the confusion matrix in Table 3. Among these, some accuracy measures more commonly used in machine learning but less in other fields included as follows:

  • F Score, the harmonic average of the precision (also positive predictive value) and recall (also sensitivity) ranging from 0 to 1 (perfect accuracy), which in our case, was 0.92;

  • Matthews correlation coefficient, a measure of the quality of binary classifications in machine learning (+ 1 representing a perfect prediction, 0 an average random prediction and − 1 inverse prediction), was 0.78;

  • The area under the precision-recall curve representing an alternative to the area under the receiver operator characteristics curve that is considered more informative for imbalanced classes. A high area under the curve, 0.88 in our case, represents both high recall and high precision.

Table 2 k-NN accuracy metrics weighted average and by class
Full size table
Table 3 Confusion matrix for the test group
Full size table

Discussion

In the previous edition of the WHO Classification of Tumors of the Pituitary Gland (2004), pituitary neuroendocrine tumors were divided into a typical adenoma, atypical adenoma, and carcinoma. Atypical adenomas were identified by an elevated mitotic index and ki-67-labeling index ≥ 3%, suggestive of aggressive clinical behavior. Using this classification, atypical adenoma incidence was relatively variable and lacking valid and reliable prognostic correlations [35]. Therefore, Trouillas and colleagues proposed a new clinicopathological classification using MRI invasiveness signs (cavernous and/or sphenoid sinus involvement), immunocytochemistry and proliferative markers, and labeling adenomas according to size, type, and grade (grade 1a, non-invasive; 1b, non-invasive and proliferative; 2a, invasive; 2b, invasive and proliferative; and 3, metastatic) [2]. Along this line, the fourth WHO classification edition (2017) emphasizes pituitary adenoma histopathological aspects and molecular genetics, abandons the term “atypical,” and strongly recommends an accurate assessment of tumor subtyping, proliferative potential, and clinical parameters such as tumor invasion for consideration of aggressive adenomas [36, 37]. However, it does not provide any hint to correlate these elements to define pituitary adenomas subtypes [38].

In this setting, artificial intelligence could contribute to better allocate individual cases in relation to aggressiveness. Machine learning applications have proved promising in recently published papers, providing insights into how predictive modeling can improve patient perioperative management. Hollon and colleagues reported an accuracy of 87% in the prediction of early postoperative outcomes in a retrospective cohort of 400 consecutive pituitary adenoma patients [39]. At the same time, Staartjes et al. demonstrated the usefulness of deep learning to preoperatively predict gross-total resection likelihood in 140 patients, reporting a 91% accuracy [40]. Furthermore, using radiomics, MRI proved accurate in predicting non-functioning pituitary adenomas subtypes [41] and cavernous sinus invasion [42].

To our knowledge, this is the first study investigating proliferative index prediction using a radiomics analysis, possibly affecting surgical approach and postoperative management. Interestingly, MRI had already proved promising in proliferative index prediction, using diffusion-weighted imaging. Indeed, a strong correlation of ADC values and ki-67 has been recently reported [43]. Our study demonstrates that data mining from non-diffusion-weighted sequences can provide similar results.

A frequent limitation of radiomic machine learning studies is applicability across different sites, varying scanner vendors and field intensity. We chose to analyze images acquired on both 1.5 and 3 T scanners by different vendors, suggesting that our results could more easily be confirmed using exams from different institutions/equipment. Obviously, this approach requires accurate pre-processing of images to reduce scanner-induced variability.

Another recurring issue in radiomics and machine learning applications is the optimal feature number choice. It depends on both sample size and algorithm employed [44]. As texture analysis often yields very large datasets, data dimensionality reduction methods are necessary to select optimal subsets for the proposed classification problem. As error distribution usually cannot be known prior to classification, it is best to test a wider range of feature set sizes derived from different selection methods. In our case, when feature selection output was represented by a ranking, a range of 4–6 parameters was used to create distinct sets. Among these, the 4-feature set ranked by direct correlation with high Ki-67 expression proved most effective in combination with k-NN.

Recently, a growing interest has been shown for deep learning applications in medical imaging [45]. Their complex technical aspects are surely more fascinating than simpler machine learning algorithms such as the k-NN we employed. On the other hand, deep learning presents its own sets of issues, such as the “black box” nature of its feature extraction and selection as well as a decision process, limiting software assumption correctness assessment and subsequent wide-scale applicability. Furthermore, data required for training such as networks is exponentially larger and computational time is also increased compared with post-processing pipelines such as the one we used. For these reasons, it would be more correct to start by using simpler, less resource-intensive machine learning methods, reserving more complex approaches in case satisfactory results are not obtainable by other means.

This study has some limitations which have to be pointed out. A further study on a more numerous population is necessary to further validate and possibly expand these results. Only T2-weighted images were used as contrast-enhanced sequences were not taken into consideration due to the presence of both gradient echo and spin echo sequences; DWI was not performed for all lesions. However, obtaining valuable data without contrast agent administration could represent an added value. Finally, while very good, feature reproducibility was only tested after their selection on a subset of patients.

Obviously, the proliferative index is only one of many aspects to be taken into consideration. However, with the proposed approach, from a T2-weighted sequence, it might be possible to obtain data concerning size, invasiveness, and proliferative index, as well as information on secretory activity [46] and on collagen content [47], useful for predicting tumor consistency. Regarding the last, a recent study by Zeynalova [48] employed a neural network on data extracted from 55 pituitary adenoma patients to assess, with good results (72.5% accuracy). It is interesting to note that only T2 images were employed, as done in our study. T2-weighted MR images also proved effective in predicting response to somatostatin analogues in patients with acromegaly and GH-secreting pituitary macroadenoma using a radiomic machine learning approach [49, 50].

For this reason, a possible future application of artificial intelligence in the study of macroadenomas could derive from consideration, in addition to data originating from advanced image analysis (intensity, texture, shape and wavelet), of clinical data, immunohistotype, proliferative indices, and invasiveness parameters (cavernous sinuses, sphenoid sinus), also evaluating other omics such as proteomics and genomics, to improve lesion classification and disease treatment choice.

Conclusion

We analyzed pituitary adenoma proliferative index class preoperative prediction, based on T2-weighted MR imaging. Our findings might aid the surgical strategy making a more accurate preoperative lesion classification and allow for a more focused and cost-effective follow-up and long-term management.